Getting started

A clear view of your agent work.

Connect your history, decide what needs attention, and set the limits that fit your work. Start with visibility. Add controls when you understand what they cover.

1. Install beside your agents

   Download and unzip your licensed copy on the Mac or Linux machine that runs Claude Code. You’ll need Python 3.11 or newer and access to its local session history.

   In a terminal inside the unpacked folder:

   ./install.sh --start-now

   Open the local address printed by the installer, normally http://127.0.0.1:4711. If a port is occupied, use the address it actually prints.

   Installation and privacy details

   The installer creates a Python environment and a launchd or systemd user service. It reads history on this machine; it does not automatically connect other computers. Keep the server on its default loopback address. The product has no login screen, so a shared remote install needs a properly authenticated proxy.

   The full instructions ship in docs/INSTALL.md. If you use a custom state directory, give the dashboard and tool-call hook the same COMMAND_CENTER_STATE_DIR or explicit COMMAND_CENTER_BREAKER_CONFIG.

2. Check what is connected

   Start at Overview. Read the data-source label before reading the numbers. If you see sample data, your own history has not been detected.

   • Sessions: confirm you recognize the projects. Open a session to inspect its recorded activity and any warnings.
   • Scheduled work: register jobs you want to check. An empty list means checks need setup; it does not mean every job is healthy.
   • Spending: use estimates to investigate usage. They are based on transcripts and a pricing table, not a Stripe or Anthropic invoice.
   Connecting scheduled work

   Command Center needs a registry describing each job and a signal that proves it ran: a last-success timestamp, log marker, status file, or expected output. It cannot infer those promises from a process name. docs/CONFIG.md includes the supported shapes and examples.

   Checks run when the workspace requests data. For continuous checks or alerts if the dashboard stops, configure an external monitor. A stopped dashboard cannot report its own failure.

3. Review the item that needs attention

   Use the attention queue to choose a session, job, or issue. Open its evidence before taking action. A high cost can belong to completed work; a missing output can mean a job is misconfigured.

   • Above a threshold: compare the recorded amount with the rule, then inspect the session and its activity.
   • Late scheduled work: check the last recorded success and expected interval before assuming the job failed.
   • Data warnings: review unreadable or unpriced records before trusting a cost total.

   Budget comparisons are observations. They are not an audit record proving a tool call was blocked or a process was stopped.

4. Set limits, then connect the hook

   Choose limits for a session, project, and day in Settings. A blank or zero limit disables that comparison. Saving a rule does not install the hook or prove that enforcement is working.

   Connect the Claude Code tool-call hook

   In your Claude Code settings.json, add this hook alongside any existing hooks. Replace both absolute paths with the location of your unpacked product. Preserve your other settings.

   {
     "hooks": {
       "PreToolUse": [{
         "hooks": [{
           "type": "command",
           "command": "/absolute/path/to/command-center/.venv/bin/python3 /absolute/path/to/command-center/control/hooks/pretooluse_breaker.py"
         }]
       }]
     }
   }

   See docs/INSTALL.md for the complete configuration. Verify the hook in a controlled Claude Code session before relying on it. The dashboard does not automatically verify that connection.

   The hook can block tool calls when recorded spend exceeds a relevant rule. It cannot guarantee a billing cap or intercept API calls made outside Claude Code.

5. Make a short review routine

   Refresh the workspace, check data coverage, review the most important open signal, and inspect the affected work. Keep a record of the diagnosis and resolution in your normal workflow.

   Stop or restart only a process you recognize and have deliberately allowlisted. Those actions require confirmation and are unavailable in the public demo.

   If something looks wrong

   Use Refresh or Retry. If the source still cannot load, the previous values are not fresh evidence. Check the configured paths and consult docs/TROUBLESHOOTING.md.

   To check the installation against bundled sample data, run .venv/bin/python3 -m ui.selfcheck. This checks the application; it does not prove your own hook or jobs are correctly configured.